Seed sweep: seed=137 (frontier verification)

[tcapelle]

Purpose

Second seed for frontier verification.

Implementation

Added before model = Transolver(...):

torch.manual_seed(137); torch.cuda.manual_seed_all(137)
import random; random.seed(137); import numpy as np; np.random.seed(137)

Results

W&B run: fern/seed-137-r18 (w40a25qs), best epoch 61, state=failed (wall-clock timeout)

Split	mae_surf_p	vs baseline
val_in_dist	19.03	+1.53
val_ood_cond	13.36	-0.94
val_tandem_transfer	38.54	+0.84
val_ood_re	27.42	-0.28
mean3	23.64	+0.44 (+1.9%)

• loss3 = 0.8542 (vs baseline 0.87 — improved)
• mean3 = 23.64 vs baseline 23.2 — slight regression (+1.9%)

Analysis

Seed variance is substantial here. Seed=137 gives loss3=0.8542 (better checkpoint metric) but mean3=23.64 (+0.44 worse). Mixed picture:

• ood_cond improves (-0.94)
• in_dist regresses (+1.53)
• tandem regresses (+0.84)

This is consistent with normal seed variance at the frontier (±1-2%). No strong signal either way — seed=137 does not unlock better performance than the default seed at this architecture.

Combined with the seed=42 run (frieren) for multi-seed statistics to confirm whether the frontier has truly plateaued.

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

0 out of 2 committers have signed the CLA.
❌ @senpai-advisor
❌ @senpai-fern
senpai-advisor, senpai-fern seem not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[tcapelle]

Closing: Phase 1 seed sweep experiment. Moving to Phase 2 with longer training runs. Seed variance results noted for reference.